context-guru plugin
A local proxy and plugin for Claude Code and Codex that trims tool output, keeps the prompt cache warm and shows the savings
- 74 ★ on the repository
- Last commit
- Apache-2.0
Broad access
Why this level
- The local proxy sees every model request of the session, including file contents and command output
- Presets `medium` and above call a model themselves and spend tokens on summaries
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/context-guruDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The local proxy sees every model request of the session, including file contents and command output
- Presets `medium` and above call a model themselves and spend tokens on summaries
This is third-party code. Review the repository files before installing.
What it does
context-guru sits as a local proxy in front of an agent session and cuts token spend two ways. First, carry less context: drop repeated tool output, collapse superseded runs and summarize before the window limit. Second, pay less for what you carry: keep the prompt cache warm and split the volatile tail off the system prompt. Cache savings are on by default, before anything trims content. Depth is set by presets: `off` changes nothing, `conservative` deterministically trims tool output with no model calls, `medium` adds a cheap model that keeps what looks relevant and a background summary, `high` uses the summary when the window is nearly full, `xhigh` runs a periodic deep pass over cooled turns and `ultrahigh` experimentally defers unused tools. `/context-guru:preset-picker` helps choose a preset. The status line shows savings for the main session and subagents, the cache clock and context fill. `/context-guru:insights` reports how much was saved and what you carry without use, and a browser dashboard keeps every request. The plugin installs in Claude Code and Codex support is experimental.
Who it is for. People who work a lot in Claude Code or Codex and want to cut token spend and latency without losing accuracy.
Good fit when
- Long sessions are expensive and the prompt cache cools down between requests
- Tool output clogs the context with repeats and dead runs
- You want to see in the status line how much was saved and how much context is left
- You want to find MCP servers and skills you never use and remove them
Not a fit when
- You want a simple skill with no proxy: the tool puts a local proxy on the request path
- You need Pi or OpenCode in enterprise mode: they are not supported yet
- You need maximum tool selection accuracy on Opus: the `ultrahigh` preset erred more often on vague tasks in the last run
Example request
Show how many tokens context-guru saved this week and find the MCP servers and skills I never useLimitations
Presets `medium` and above spend tokens themselves: they call a cheap model and prepare summaries. The `ultrahigh` preset is experimental: in the last run on Opus 5.5 it picked the right tool in 66 of 80 cases against 72 of 80 with no deferral, while on Sonnet 5.5 it matched `xhigh`. In Claude Code a preset change applies at the next proxy start, while Codex restarts its own proxy. Codex does not check for binary updates automatically, so they are installed by hand. Pi and OpenCode are marked experimental or not yet available.
How to disable. Pick the `off` preset through `/context-guru:preset-picker` or remove the plugin with `/plugin uninstall context-guru@context-guru`. In Codex use `codex plugin marketplace upgrade` and the CLI plugin management.
FAQ
Which preset should I start with?
`conservative`: it trims tool output deterministically and makes no model calls. The other presets from `medium` upward spend tokens on filtering and summaries.
What does /context-guru:insights show?
Over seven days: how much you saved, what your cache costs and which context you carry and never use. The data comes from your real sessions.
Is an API key needed?
The marketplace entry for the plugin says it works with no API key on a Pro or Max subscription. There is a separate setup for connecting to an organization.
Does it work with Codex?
Experimentally: it installs with `codex plugin marketplace add` and updates with `codex plugin marketplace upgrade context-guru`, followed by a Codex restart.
What does ultrahigh do?
It sends the model only the names of unused tools and loads the one it needs through search. The preset is experimental and is opt-in.
Discussion
No questions or reviews yet.
Related
PluginsThe context-optimizer plugin shows Claude session cost, trims context and hands file reads and tests to cheap Haiku agents.
A plugin that cuts token use in long sessions: prunes stale tool output, picks a subagent model and narrows the skill list.
An installer for a set of hooks, commands and plugins that compress tool output and save tokens in Claude Code.
Microsoft Research and UC Berkeley plugin that cuts the cost of Claude Code, Codex and Copilot CLI sessions using earlier sessions
A Rust program that shrinks the command output an agent reads: drops build and test noise, folds repeated file reads and redacts credentials
A plugin for long Claude Code runs: context, cache and progress at a glance, with handoff to a fresh context before it fills.
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| context-guru (this card) | no | 74 | ||
| Token Optimizer | no | 2,964 | a plugin that cuts tokens and keeps work across sessions | |
| LeanToken | no | not required | 158 | local code search that keeps context within a budget |
README in short
The README presents the tool as context engineering for agents: less context and a cheaper cache. Install choices for Claude Code, Codex, Pi and OpenCode, the status line and commands, a preset table, the experimental mode and the update procedure follow.